Which is better, DeepSeek-R1 or DeepSeek-V3?

Across 63 shared benchmarks tracked on Vector Wire, DeepSeek-R1 leads on 52 and DeepSeek-V3 leads on 11 — DeepSeek-R1 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-R1 vs DeepSeek-V3

DeepSeekvsDeepSeek63 shared benchmarks52–11 head-to-head

Benchmark	DeepSeek-R1	DeepSeek-V3
AA Agentic Index	3.1	1.6
AA Intelligence	18.6	15.4
AA-LCR	56	41.3
AA-Omniscience	-31.3	-37.6
aider_polyglot	71.6	55.1
AIME 2024	91.4	59.4
AIME 2025	87.5	51.3
AIR-Bench 2024	52.9	40.8
AlpacaEval2.0 (LC-winrate)	87.6	70
anthropic_red_team	99.1	97.1
Arena-Hard (GPT-4-1106 judge)	92.3	85.5
ArenaHard (GPT-4-1106)	92.3	85.5
Artificial Analysis Coding Index	24.6	23
bbq	96.6	96.7
C-Eval	91.8	90.1
Chinese SimpleQA (C-SimpleQA)	63.7	68
CLUEWSC	92.8	90.9
CNMO 2024	78.8	74.7
Codeforces	2029	1134
Codeforces (Percentile)	96.3	58.7
Codeforces (Rating)	2029	1134
critpt	0.6	0
DROP (3-shot F1)	92.2	91.6
FRAMES (Acc.)	83	73.3
gdpval	1.5	0
GPQA Diamond	81	68.4
harmbench	54.6	49.7
HLE	17.7	5.2
HMMT 2025	79.4	27.5
hmmt_feb_2025	76.7	29.2
IFBench	39	41
ifeval	83.3	87.3
LiveCodeBench	84.4	49.6
livecodebench_easy	99.2	83.3
livecodebench_hard	63.6	14.2
livecodebench_medium	90.9	53.5
livecodebench_pass1cot	65.9	40.5
longbench_v2	58.3	48.7
MATH	97.3	91.2
math_500_em	98	94
mmlu	92.9	89.4
mmlu_prox	75.5	70.5
mmlu_redux	93.4	90.5
MMLU-Pro	85	81.2
multichallenge	45	31.4
OmniScience Accuracy	30.7	25.4
OmniScience Non-Hallucination	10.5	23.3
scicode	35.7	35.8
simple_safety_tests	98.3	95.3
simplebench	40.8	27.2
simpleqa	92.3	27.7
SWE-bench Verified	57.6	42
TauBench V3 - Banking	6.4	4.7
Terminal-Bench 2.1	19.1	16.9
Terminal-Bench Hard	6.1	15.2
vectara_answer_rate	97	97.5
vectara_avg_summary_length	93.5	81.7
vectara_factual_consistency	88.7	93.9
vectara_hallucination_rate ↓	11.3	6.1
xstest	98.8	97.1
ZebraLogic	95.1	84
τ²-Bench Telecom (AA run)	11.4	47.1
τ³-Bench	6.4	4.7

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.